将 tf.contrib.layers.batch_norm 迁移到 Tensorflow 2.0

Question

I'm migrating a TensorFlow code to Tensorflow 2.1.0.

Here is the original code:

conv = tf.layers.conv2d(inputs, out_channels, kernel_size=3, padding='SAME')
conv = tf.contrib.layers.batch_norm(conv, updates_collections=None, decay=0.99, scale=True, center=True)
conv = tf.nn.relu(conv)
conv = tf.contrib.layers.max_pool2d(conv, 2)

And this is what I've done:

conv1 = Conv2D(out_channels, (3, 3), activation='relu', padding='same', data_format='channels_last', name=name)(inputs)
conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last")(conv1)
#conv = tf.contrib.layers.batch_norm(conv, updates_collections=None, decay=0.99, scale=True, center=True)
pool1 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last")(conv1)

My problem is that I don't know what to do with tf.contrib.layers.batch_norm .

How can I migrate tf.contrib.layers.batch_norm to Tensorflow 2.x?

UPDATE:
Using the comment suggestion, I think I have migrated correctly:

conv1 = BatchNormalization(momentum=0.99, scale=True, center=True)(conv1)

But I'm not sure if decay is like momentum and I don't know how to set updates_collections in the BatchNormalization method.

Answer 1

I encountered this problem when working with a trained model that I was going to fine tune. Just replacing tf.contrib.layers.batch_norm with tf.keras.layers.BatchNormalization like OP did gave me an error whose fix is described below.

The old code looked like this:

tf.contrib.layers.batch_norm(
    tensor,
    scale=True,
    center=True,
    is_training=self.use_batch_statistics,
    trainable=True,
    data_format=self._data_format,
    updates_collections=None,
)

and the updated working code looks like this:

tf.keras.layers.BatchNormalization(
    name="BatchNorm",
    scale=True,
    center=True,
    trainable=True,
)(tensor)

I'm unsure if all the keyword arguments I removed are going to be a problem but everything seems to work. Note the name="BatchNorm" argument. The layers use a different naming schema so I had to use the inspect_checkpoint.py tool to look at the model and find the layer names which happened to be BatchNorm .